Capacity Constrained Optimal Trade Acceptance in Corporate Bond Market Making

Abstract

We study how a capacity-constrained market maker in the investment-grade corporate bond market should decide whether to accept or reject arriving trade opportunities. We formulate the problem as an event-time constrained Markov decision process calibrated to TRACE and FISD data, and compare simple benchmark policies, shadow-pricing rules, and a reinforcement learning policy trained using masked PPO. Our results show that trade acceptance is fundamentally a dynamic inventory-allocation problem: policies that ignore capacity usage, inventory persistence, or cross-cell heterogeneity perform substantially worse, while PPO achieves the strongest overall performance by conditioning on the full inventory state. We extract a practical heuristic from the full-state learned policy and interpret its implications for optimal trade acceptance in the corporate investment-grade bond market.